Re: RXE status in the upstream rping using rxe
From: Bob Pearson <hidden>
Date: 2021-08-20 22:09:56
On 8/20/21 4:40 PM, Bob Pearson wrote:
On 8/20/21 2:42 AM, yangx.jy@fujitsu.com wrote:quoted
On 2021/8/20 11:31, Zhu Yanjun wrote:quoted
Latest kernel + latest rdma-coOnre< ------rping----> 5.10.y stable + latest rdma-core Latest kernel + latest rdma-core< ------rping----> 5.11.y stable + latest rdma-core Latest kernel + latest rdma-core< ------rping----> 5.12.y stable + latest rdma-core Latest kernel + latest rdma-core< ------rping----> 5.13.y stable + latest rdma-core The above works well.Hi Yanjun, Sorry, I don't know why you cannot reproduce the bug. Did you see the similar bug reported by Olga Kornievskaia? https://www.spinics.net/lists/linux-rdma/msg104358.html https://www.spinics.net/lists/linux-rdma/msg104359.html https://www.spinics.net/lists/linux-rdma/msg104360.html Best Regards, Xiao Yangquoted
Zhu YanjunThere is some interest in the current status of rping on rxe. I have looked at several configurations and tested the following test cases: 1. The python test suite in rdma-core 2. ib_xxx_bw and ib_xxx_bw -R for RC 3. rping Between the following node configurations. A. 5.11.0 (ubuntu 21.04 OOB) + rdma-core 33.1 (ubuntu 21.04 OOB) B. 5.11.0 + current rdma-core + "Provider/rxe:Set the correct value of resid for inline data" (a.k.a rdma-core+) C. 5.14.0-rc1+ (for-next current) + 5 recent bug fixes (a.k.a. for-next+) RDMA/rxe:Fix bug in get srq wqe in rxe_resp.c.patch RDMA/rxe:Fix bug in rxe_net.c.patch RDMA/rxe:Add memory barriers to kernel queues.patch RDMA/rxe:Fix memory allocation while locked.patch RDMA/rxe:Zero out index member of struct rxe_queue.patch + rdma-core+ D. for-next+ + rdma-core (33.1) Results: 1. A N/A 1. B no errors, some skips 1. C no errors, some skips 1. D N/A (n.b. requires adding IPV6 address == gid[0] by hand) 2. [A-D] -> [A-D] all pass 3. A -> A, C -> C, D -> D all pass, all other combinations fail (RDMA_resolve_route: No such device. Not yet sure cause of failures but looking into it.) In theory these should all work but rdmacm is more sensitive to configuration than verbs. Bob
Found the problem (thank you google) If you run both server$ rping -s -a nn.nn.nn.nn client$ rping -c -a nn.nn.nn.nn now all tests pass for rping as well. Bob